Continuous-variable photonic quantum extreme learning machines for fast collider-data selection

TL;DR

This work introduces continuous-variable quantum extreme learning machines (CV-QELMs) as ultra-fast front-ends for collider data, leveraging displacement-based input encoding on $M$ qumodes, a fixed Gaussian quantum substrate, and Gaussian measurements to produce a high-dimensional feature map of dimension $R$ that feeds a linear readout trained analytically. By comparing against parameter-matched MLP baselines on top jet tagging and Higgs identification tasks, the study shows CV-QELMs can outperform small networks and rival larger ones while maintaining fixed, nanosecond-scale inference latency and minimal training overhead. The results highlight the practicality of Gaussian photonic random features for real-time data selection near detectors and point to hardware co-design paths for future trigger integration. Overall, CV-QELMs offer a robust, low-power, fast alternative for online collider data processing with potential deployment close to the detector.

Abstract

We study continuous-variable photonic quantum extreme learning machines as fast, low-overhead front-ends for collider data processing. Data is encoded in photonic modes through quadrature displacements and propagated through a fixed-time Gaussian quantum substrate. The final readout occurs through Gaussian-compatible measurements to produce a high-dimensional random feature map. Only a linear classifier is trained, using a single linear solve, so retraining is fast, and the optical path and detector response set the analytical and inference latency. We evaluate this architecture on two representative classification tasks, top-jet tagging and Higgs-boson identification, with parameter-matched multi-layer perceptron (MLP) baselines. Using standard public datasets and identical train, validation, and test splits, the photonic Quantum Extreme Learning Machine (QELM) outperforms an MLP with two hidden units for all considered training sizes, and matches or exceeds an MLP with ten hidden units at large sample sizes, while training only the linear readout. These results indicate that Gaussian photonic extreme-learning machines can provide compact and expressive random features at fixed latency. The combination of deterministic timing, rapid retraining, low optical power, and room temperature operation makes photonic QELMs a credible building block for online data selection and even first-stage trigger integration at future collider experiments.

Figures (8)

• Figure 1: Visualisation of the QELM. The features of the data get encoded onto qumodes via displacement operations. A fixed-time Gaussian quantum substrate $\hat{U}_G$ followed by Gaussian-compatible measurements creates a random, high-dimensional feature map, which is used, together with the original data, in a linear classifier.
• Figure 2: Schematic of the photonic substrate for the quantum extreme learning machine (QELM). Classical data are first embedded into optical modes via displacement operations before being propagated through a Gaussian circuit composed of cascading controlled-addition ($\hat{C}_X$) gates. Homodyne measurements are then performed on the output qumodes to generate the feature vector, which serves as input for downstream classical classification.
• Figure 3: Scaling of test accuracy with training set size for the QCD versus Top jet classification task with $F=16$ features pierini_2020_3602260. Mean test accuracy is shown for the CV-QELM with (left) PNR and (right) homodyne measurements using both logistic and ridge regression readouts, compared against an MLP baseline with $H=2$. Error bars denote the standard deviation across repeated runs. The CV-QELM consistently outperforms the MLP across all training sizes, maintaining high accuracy and low variance even in the low-statistics regime.
• Figure 4: QCD versus Top jet classification with $F=16$ features and $n_{\mathrm{samples}}=10^5$ from the hls4ml dataset pierini_2020_3602260 using logistic regression on the CV-QELM outputs. Accuracy distributions are shown for the CV-QELM with (left) PNR and (right) homodyne measurements, compared against MLP baselines with $H=2$ and $H=10$. The CV-QELM achieves higher mean accuracy and substantially lower variance than both MLP baselines, demonstrating the stability and expressiveness of the Gaussian photonic random feature representation.
• Figure 5: QCD versus Top jet classification with $F=16$ features and $n_{\mathrm{samples}}=10^5$ from the hls4ml dataset pierini_2020_3602260 using ridge regression on the CV-QELM outputs. Accuracy distributions are shown for the CV-QELM with (left) PNR and (right) homodyne measurements, compared against MLP baselines with $H=2$ and $H=10$. The CV-QELM achieves higher mean accuracy and markedly lower variance than both MLP baselines, highlighting the robustness and efficiency of the analytic ridge regression readout in combination with Gaussian photonic random features.